Evidence before
confidence.
Train a model. Set your criteria. Challenge its performance.
A practical bridge between laboratory science and machine learning.
Build your experiment
01 / Validation1,200 synthetic observations. Stratified 60% training / 20% validation / 20% test. Preprocessing is fitted on training data only.
Header: id,label,score. Labels 0 or 1; scores 0–1. Positive means “QC failure”. Evaluation only; imported predictions are not trained here. Upload only non-sensitive data.
Predict failure when score ≥ threshold. A lower threshold catches more failures but may reject more acceptable runs.
Can an assay run be flagged
for quality review?
Learn a binary classifier from three simulated instrument features: signal drift, replicate variability and background signal. The reference label marks whether a run failed quality control.
Receiver operating characteristic
AUC measures ranking across thresholds. It does not establish calibration or suitability for deployment.
Performance, with context
Validation subsetConfusion matrix
Rows: reference label. Columns: predicted label.
| Flagged failure | Acceptable | |
|---|---|---|
| Actual failure | ||
| Actual acceptable |
False negative = missed QC failure. False positive = unnecessary rejection of an acceptable run.
Calibration check
Mean predicted score vs observed failure frequency, in five probability bins. Small bins are unstable; the Brier score measures overall probability error.
Error review
| Case | Reference | Score | Prediction | Outcome |
|---|
Model evidence
A model you can inspect
Coefficients apply to training-standardised features. They describe association in this synthetic dataset, not causation.
Training loss
Make the evaluation defensible.
01 / Define the claim
The positive class is a simulated assay QC failure. This app does not diagnose disease or evaluate biological threat capabilities.
02 / Separate the evidence
Fit scaling and logistic regression on training data. Select a threshold using validation data. Lock it before revealing the held-out test set.
03 / Record uncertainty
Sensitivity and specificity include 95% Wilson intervals. An absent denominator is reported as N/A. Acceptance can use point estimates or interval lower bounds.
04 / State the limits
Synthetic outcomes are sampled from a logistic process. Good performance is expected and does not establish real-world validity. Repeated experiments after seeing test results are exploratory.
Implementation: browser-native binary logistic regression, full-batch gradient descent, learning rate 0.12 and optional L2 penalty (intercept excluded). The loss plot shows unpenalised training cross-entropy. ROC AUC uses tied-score grouping and trapezoidal integration.
CSV imports are evaluated as an external scored dataset. The app cannot verify how the external model was trained or whether leakage occurred. No live LLM connection is configured. Data stays in the current browser tab and is cleared on reload.
Read about data leakage ↗ Logistic regression reference ↗